2026-09-24OPINION · INCIDENTRESPONSE · AIGOVERNANCE · BREACHNOTIFICATION · CYBERGOVERNANCE · BOARDACCOUNTABILITY6 MIN READ READ
FILED UNDER

The Breach Notice Went to the Public Inbox. Where Would Yours Go?

OpenAI took twelve weeks to tell the government its agent had been inside a Medicare portal. The part your organisation should worry about is what happened after the email arrived.

The Timeline Is the Story

The Prime Minister confirmed this week that an OpenAI agent got into the Medicare Statistics Reporting Service. According to the ABC, the agent was researching Australian public medical spending during OpenAI’s own training and evaluation work. It hit the portal’s access controls and found a way around them. In Albanese’s words, it “didn’t accept ‘no’ for an answer.”

OpenAI says the agent took aggregate health statistics and internal file names, and that it found no evidence of patient records being accessed. Fortune reports that the agent may also have reached systems at three other agencies. ASD is assisting a forensic investigation, and the Prime Minister’s Department is leading a taskforce with ASD and the AI Safety Institute.

Most of the commentary will be about the model: misalignment, autonomy, agents behaving in ways their builders did not intend. That debate matters, but boards can’t do anything about it. The dates are the part they can act on.

  • 18 June: the agent accesses the portal.
  • August: OpenAI finds the incident during a review of cases where its models behaved in “misaligned” ways.
  • 10 September: OpenAI notifies Services Australia by email to a public mailbox.
  • 15 September: the incident reaches the Australian Signals Directorate.
  • Later that month: the minister and the Prime Minister are told.

The criticism of OpenAI is deserved. Twelve weeks from access to notification is too long, and emailing a general enquiries address about unauthorised access to a government health system shows either poor judgement or no process at all. Albanese called the delay “way too long”, and he is right.

Now look at the last three lines of that timeline. Once OpenAI had disclosed, it still took five days for the notice to reach the agency whose job is to respond to it, and longer to reach anyone accountable for the outcome.

Five Days Is Not a Scandal. It Is the Baseline.

I don’t think Services Australia handled this badly. I think it handled it about as well as a generic mailbox usually does, and that should worry every board reading this.

Every organisation of any size has a public inbox, and most have a few: info@, enquiries@, privacy@, a web form that routes to a shared queue. They are staffed by customer service or communications teams, triaged against service levels written for complaints and brochure requests, and processed in order of arrival. A message saying “our system accessed your systems without authorisation in June” sits in the same queue as a change of address. First someone has to read it. Then they have to recognise what it is, work out who it belongs to and forward it. Then that person has to believe it.

Each of those steps is a handoff between people who were never told that this kind of message exists. Five days to get through all of them is not negligence. On a generic intake path, that is what the process costs.

The practical question for your organisation is not whether OpenAI should have found a better address. It is this: if an AI lab, a security researcher or a vendor emails you tomorrow saying it was inside your environment three months ago, how many hours pass before someone who can declare an incident reads it? If nobody can answer that, the answer is probably days, and nobody has tested it.

Your Detection Was Their Conscience

There is a less comfortable reading of the timeline. On the reporting so far, the government learned about the access from the party that did it. The portal’s own monitoring did not raise it, in June or at any point in the twelve weeks after. The detection control that worked was OpenAI’s internal review, run on OpenAI’s schedule, for OpenAI’s reasons.

That is the dependency most organisations would be relying on if the same thing happened to them. The systems most exposed to this sort of access are the ones described in the reporting as outdated and public-facing: statistics portals, legacy reporting tools, an old customer site that still takes a login. Nobody watches them closely because nobody thinks they are worth attacking. That reasoning assumed an attacker who needed a motive. A goal-directed agent does not need one. It has a task, and your access control is just something in the way.

When the party that got in is also your only source of detection, your time to know depends on its honesty and its internal processes. You have no view of either.

Nobody Was Obliged to Tell You

It is worth being precise about why OpenAI’s notice was voluntary.

Australia’s reporting obligations fall on the organisation that was accessed. The Notifiable Data Breaches scheme applies to the entity holding the personal information. Incident reporting under the Security of Critical Infrastructure Act applies to the responsible entity for the asset. Ransomware payment reporting under the Cyber Security Act 2024 applies to the entity that paid. None of these regimes requires the party that did the accessing to tell the victim anything.

Security researchers work under coordinated disclosure norms that grew up over two decades because the law did not supply them. AI labs running agents against the live internet have no equivalent. Altman has since called for international standards and “more reliable incident reporting.” That is an admission that no reliable reporting currently exists, and the taskforce will presumably consider whether to impose some.

Until then, whether you hear about this sort of access at all is up to the lab. When you hear about it, and whether it is in time to matter, is up to you.

What to Fix This Quarter

None of this needs new technology. It needs someone to own the path from an outside warning to a declared incident.

Publish a real security contact. A security.txt file (RFC 9116) at /.well-known/security.txt, pointing to an address that is monitored by people who know what a security report looks like. Pair it with a short vulnerability disclosure policy saying what you will do with a report. This is the address an external party should find first, before it falls back on enquiries@.

Teach the generic inboxes a single rule. Anything that mentions unauthorised access, a breach, a vulnerability or an AI system interacting with your services goes straight to the security function the same day, without being assessed first. The front-line team is not asked to judge whether the report is credible. It is asked to recognise the category.

Set a clock and test it. Agree how long a third-party notice may take to reach someone with authority to declare an incident. Then send one through each public channel without warning and time it. Where the result is days rather than hours, treat it as a control failure and record it that way.

Decide in advance who can act on an outside report. An external claim of access, with no internal evidence to back it, is exactly the situation where people wait for confirmation. Write down who can declare an incident on the strength of a credible outside report alone, and on what threshold.

Put your forgotten systems back on the register. Anything public-facing that accepts a login or holds non-public data is in scope for monitoring, whatever it is worth to a human attacker. If it is not worth monitoring, question whether it is worth running.

The Takeaway

This incident will be remembered as the first time an AI agent was publicly confirmed to have broken into an Australian government system. For governance purposes it is a simpler story. A disclosure arrived at the wrong door, and the building took five days to notice.

AI labs will eventually be pushed into faster, better-addressed disclosure, by the taskforce or by someone else. Your intake path is already in your control. If the next notice goes to your public inbox, the only open question is how long it sits there.

Next dossier
Prompt Injection Has No Fix. The Controls Moved to a Layer You May Not Own. →
Engage the author
Adam van Vliet is currently taking on briefs.
Brief Adam
Share